Anthropic Says Claude Now Leads Quarter of Internal AI Research

Anthropic disclosed on Thursday that its Claude artificial intelligence model now independently leads 26 percent of the company's internal research and development tasks. This marks a sharp escalation from March, when the model contributed to only 1 percent of such workloads, according to metrics tracked by the non-profit research organization Epoch AI. The company reported that while Claude remains under human supervision, it can execute complex operations from beginning to end based on high-level prompts.

To measure this progression, Anthropic utilizes an automation rating scale developed by Epoch AI that tracks Automation Level, or AL, running from AL0 with no AI involvement to AL5 where systems operate fully autonomously. As of August, Anthropic stated that approximately 90 percent of its research activities involved collaboration with Claude, a designation where the model handles substantial portions of work under close human direction. During that same month, the company deployed roughly 30,000 AI agents across its internal platforms to execute engineering and research tasks.

Operating thousands of autonomous agents requires stringent safety protocols to prevent unexpected behaviors or recursive self-improvement risks. Anthropic screens every action executed by its internal agents, noting that it blocked approximately one in every 47,000 decisions made by the systems during August. To maintain transparency, the company published these operational metrics to encourage competing artificial intelligence laboratories to adopt standardized benchmarks for tracking automated research development.

Resource allocation data released by the company showed that 6 percent of computing power dedicated to research during a sample week in July went toward safety evaluations. That figure doubled to 12 percent for research workloads handled directly by the AI agents themselves. Anthropic clarified that these measurements remain conservative because computing capacity utilized for both capability and safety tasks was categorized entirely under capability work.

The disclosures emerge amid broader technology industry discussions regarding recursive self-improvement, where advanced models could eventually develop their successors with minimal human intervention. Researchers have repeatedly cautioned that as agent autonomy expands, system behaviors can diverge from initial developer intentions. In response to these governance challenges, Anthropic has committed to embedding external third-party evaluators within its organization to monitor internal safety frameworks. The announcement follows similar transparency initiatives from rival developer OpenAI, which recently established routine reporting schedules for unauthorized or unexpected model behaviors as regulatory scrutiny over artificial intelligence safety intensifies.